Neonatal Sepsis Prediction Using AI-Based Vital Sign Analysis in NICUs: A Multicenter Retrospective Study.

Abstract

Neonatal sepsis remains one of the leading causes of morbidity and mortality in intensive care units, yet early diagnosis continues to be a major clinical challenge due to nonspecific symptoms and rapidly evolving physiological instability. This study develops and evaluates an AI-based vital-sign prediction framework capable of detecting early signatures of sepsis using continuous physiological monitoring data from multiple NICUs. A multicenter retrospective dataset comprising heart rate, respiratory rate, oxygen saturation, temperature, blood pressure, and derived variability metrics was analyzed across three tertiary hospitals. After preprocessing, imputation, and artifact suppression, machine learning and deep learning models including Random Forest, Gradient Boosting, and a hybrid LSTM network were trained to capture nonlinear temporal patterns indicative of sepsis onset. The proposed LSTM-based model demonstrated strong predictive ability, identifying sepsis up to 6–12 hours before clinical diagnosis with improved sensitivity and reduced false alarms compared to conventional scoring systems. Key physiological precursors included abnormal heart-rate variability, intermittent desaturations, rising temperature instability, and increased respiratory fluctuations. The findings confirm that AI-guided prediction can enhance real-time surveillance, support timely clinical interventions, and reduce progression to severe outcomes. This multicenter analysis underscores the potential of intelligent vital-sign analytics as a scalable decision-support tool in modern NICU environments.